What happened
AWS announced that NVIDIA Nemotron 3.5 Super VL is available in SageMaker JumpStart, deployable through SageMaker Studio, the SageMaker Python SDK or a Hugging Face model page. It is the vision-language model in NVIDIA’s Nemotron 3.5 family: it takes images, video and text as inputs and returns text. The model combines a 120-billion-parameter hybrid Mixture-of-Experts language backbone, with 12 billion active parameters per token, and a C-RADIO vision encoder. AWS says it supports context lengths up to one million tokens, configurable reasoning and a built-in Multi-Token Prediction head; however, the JumpStart serving configuration does not enable speculative decoding.
JumpStart offers NVFP4 and BF16 checkpoints. NVFP4 is quantized from BF16 with quantization-aware distillation, uses 4-bit weights and activations, and is positioned for optimized inference. Its checkpoint is about 75 GB and runs on G7 instances: the default ml.g7.24xlarge with four NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, or ml.g7.48xlarge with eight. BF16 is about 249 GB and runs on ml.p5en.48xlarge, with eight NVIDIA H200 GPUs; AWS describes it as a reference for customization such as supervised fine-tuning, reinforcement learning, distillation and domain adaptation. The license listed for both is OpenMDW License Agreement 1.1. Neither endpoint variant supports LoRA adapter serving.
AWS positions the model for document intelligence, GUI understanding for computer-use agents, video question answering and multimodal agent workflows. Its examples use a vLLM-powered chat-completions API, with tool calling, JSON-schema output and per-request reasoning configuration. Media must be supplied inline as base64 data URLs because the endpoint is network-isolated and cannot retrieve URLs. The endpoint accepts images, not raw PDFs; a 1,536-by-1,024 page image uses about 1,536 context tokens. The JumpStart configuration accepts one image and one video per request. Serving profiles set different context limits: 2,275 tokens for generate, 13,650 for summarize and 15,600 for modify, rather than exposing the model’s full million-token context in these configurations.
AWS measured performance using 250 input tokens, 1,500 output tokens and five concurrent streaming requests. Mean output throughput per request was 63.07 tokens per second on NVFP4 ml.g7.24xlarge, 64.01 on NVFP4 ml.g7.48xlarge, and 140.85 on BF16 ml.p5en.48xlarge; mean time to first token was 0.45, 0.51 and 0.53 seconds, respectively. These are different GPU systems, not a controlled precision comparison. AWS says deploying on the default G7 instance takes about ten minutes. New accounts may start with zero G7 quota and need to request an increase.
What the archive adds
New Runtime covered Nemotron 3.5 Super VL as an open 120B multimodal model on October 21, 2026, and noted its possible role in specialized agent workflows. This is a new deployment-access fact on a known model story, not a new model release.
Trend signal
This adds a deployment option for a model intended to combine perception, reasoning and tool use, but it does not yet establish an interoperability layer or a new agent runtime boundary. This does not yet fit a pattern New Runtime tracks.
Why now
AWS’s announcement describes the JumpStart availability as current and supplies deployment configurations and examples. It does not explain a broader market or technical trigger for the timing.
Why it matters
The practical choice is not simply between a smaller and larger model. Teams can test a quantized checkpoint on G7 hardware or use BF16 reference weights on H200s for customization, but they must account for endpoint capacity, quota, context profiles and media preprocessing. The one-million-token model context should not be treated as the effective context available in the published JumpStart serving profiles: those top out at 15,600 tokens. Likewise, the reported throughput figures are useful for planning a test, not a direct precision benchmark, because the variants use different hardware. For builders, this turns evaluation into an operational decision: check whether the supported instance and license fit the workload, then benchmark with representative image or video payloads and the profile your application will actually use.
Limits and uncertainty
The performance figures are AWS measurements under one stated workload profile, not independent results or a guarantee for a team’s prompt lengths, concurrency or media mix. AWS explicitly notes that hardware differs between the BF16 and NVFP4 results. The excerpted announcement describes benchmark scores for selected vision-language tasks, but those scores are model-card results, not evidence of production accuracy for any particular document, interface or video workflow. The announcement does not establish endpoint cost-effectiveness for a specific workload. Teams should validate quality, latency, request limits and operating cost against their own data before deployment.
Skill to build now
Build a small evaluation set from the inputs your workflow actually handles: representative document pages, application screenshots and short video clips. Render PDFs into images, record each image’s context use, and test the JumpStart profile and reasoning mode you intend to run. Include failure checks for one-image-per-request handling and base64 payload limits before connecting the endpoint to an agent.
Reader takeaway
SageMaker JumpStart makes Nemotron 3.5 Super VL easier to try on AWS, but the managed endpoint is not the same as exposing the model’s full advertised context or capability set. Choose the checkpoint and GPU for the job, then test the actual serving limits and media pipeline before committing a workflow to it.